Beyond Prediction Gain: LASR as a Metric for Action-Grounded Latents in World Models
Better prediction alone does not show that a world model’s latent variables encode action.
The authors introduce LASR to measure whether learned latents help predict action-related targets beyond what current state and non-action shortcuts explain. GeoLASR summarizes typical support; WorstLASR captures the weakest case. In tests on recorded robot manipulation data and a matched multi-task panel, the metric detected known action information and rejected representative shortcuts. A lightweight adapter produced better point estimates, with uncertainty assessed through paired comparisons.
ArXiv · AI/CL/LG's note
The authors introduce LASR to measure whether learned latents help predict action-related targets beyond what current state and non-action shortcuts explain. GeoLASR summarizes typical support; WorstLASR captures the weakest case. In tests on recorded robot manipulation data and a matched multi-task panel, the metric detected known action information and rejected representative shortcuts. A lightweight adapter produced better point estimates, with uncertainty assessed through paired comparisons.
ArXiv · AI/CL/LG's note
score 4